github/ROCm
1
1
Fork 0
mirror of https://github.com/ROCm/ROCm.git synced 2026-01-08 06:13:59 -05:00
Code Issues Packages Projects Releases Wiki Activity

Update precision support page with missing libraries and RDNA2 and CDNA4 support

Browse Source
This commit is contained in:
Istvan Kiss
2025-08-28 17:09:34 +02:00
committed by GitHub
parent 04beef8773
commit d476d09aff
6 changed files with 883 additions and 337 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
.wordlist.txt
View File

@@ -96,6 +96,7 @@ DIMM
DKMS
DL
DMA
DOMContentLoaded
DNN
DNNL
DPM
@@ -531,6 +532,7 @@ ZenDNN
accuracies
activations
addr
addEventListener
ade
ai
alloc
@@ -564,6 +566,7 @@ boson
bosons
br
BrainFloat
btn
buildable
bursty
bzip
@@ -575,6 +578,7 @@ centric
changelog
checkpointing
chiplet
classList
cmake
cmd
coalescable
@@ -586,6 +590,7 @@ composable
concretization
config
conformant
const
constructible
convolutional
convolves
@@ -649,6 +654,7 @@ exascale
executables
ffmpeg
filesystem
forEach
fortran
fp
framebuffer
@@ -657,6 +663,7 @@ galb
gcc
gdb
gemm
getAttribute
gfortran
gfx
githooks
@@ -808,6 +815,8 @@ recommenders
quantile
quantizer
quasirandom
querySelector
querySelectorAll
queueing
qwen
radeon
@@ -870,9 +879,11 @@ scalability
scalable
scipy
seealso
selectedTag
sendmsg
seqs
serializers
setAttribute
sglang
shader
sharding
@@ -899,6 +910,7 @@ symlink
symlinks
sys
tabindex
targetContainer
td
tensorfloat
th

BIN
docs/data/about/compatibility/floating-point-data-types.png
View File

Binary file not shown.
Side by Side Swipe Overlay

Before

Width:  |  Height:  |  Size: 81 KiB

After

Width:  |  Height:  |  Size: 114 KiB

391
docs/data/reference/precision-support/precision-support.yaml Normal file
View File

@@ -0,0 +1,391 @@
# rocm-library-support.yaml
library_groups:
- group: "ML & Computer Vision"
tag: "ml-cv"
libraries:
- name: "Composable Kernel"
tag: "composable-kernel"
doc_link: "composable_kernel:reference/Composable_Kernel_supported_scalar_types"
data_types:
- type: "int8"
support: "✅"
- type: "int32"
support: "✅"
- type: "float4"
support: "✅"
- type: "float6 (E2M3)"
support: "✅"
- type: "float6 (E3M2)"
support: "✅"
- type: "float8 (E4M3)"
support: "✅"
- type: "float8 (E5M2)"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "MIGraphX"
tag: "migraphx"
doc_link: "amdmigraphx:reference/cpp"
data_types:
- type: "int8"
support: "⚠️"
- type: "int16"
support: "✅"
- type: "int32"
support: "✅"
- type: "int64"
support: "✅"
- type: "float8 (E4M3)"
support: "✅"
- type: "float8 (E5M2)"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "MIOpen"
tag: "miopen"
doc_link: "miopen:reference/datatypes"
data_types:
- type: "int8"
support: "⚠️"
- type: "int32"
support: "⚠️"
- type: "float8 (E4M3)"
support: "⚠️"
- type: "float8 (E5M2)"
support: "⚠️"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "⚠️"
- type: "float32"
support: "✅"
- type: "float64"
support: "⚠️"
- group: "Communication"
tag: "communication"
libraries:
- name: "RCCL"
tag: "rccl"
doc_link: "rccl:api-reference/library-specification"
data_types:
- type: "int8"
support: "✅"
- type: "int32"
support: "✅"
- type: "int64"
support: "✅"
- type: "float8 (E4M3)"
support: "✅"
- type: "float8 (E5M2)"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- group: "Math Libraries"
tag: "math-libs"
libraries:
- name: "hipBLAS"
tag: "hipblas"
doc_link: "hipblas:reference/data-type-support"
data_types:
- type: "float16"
support: "⚠️"
- type: "bfloat16"
support: "⚠️"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "hipBLASLt"
tag: "hipblaslt"
doc_link: "hipblaslt:reference/data-type-support"
data_types:
- type: "int8"
support: "✅"
- type: "float4"
support: "✅"
- type: "float6 (E2M3)"
support: "✅"
- type: "float6 (E3M2)"
support: "✅"
- type: "float8 (E4M3)"
support: "✅"
- type: "float8 (E5M2)"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- name: "hipFFT"
tag: "hipfft"
doc_link: "hipfft:reference/fft-api-usage"
data_types:
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "hipRAND"
tag: "hiprand"
doc_link: "hiprand:api-reference/data-type-support"
data_types:
- type: "int8"
support: "Output only"
- type: "int16"
support: "Output only"
- type: "int32"
support: "Output only"
- type: "int64"
support: "Output only"
- type: "float16"
support: "Output only"
- type: "float32"
support: "Output only"
- type: "float64"
support: "Output only"
- name: "hipSOLVER"
tag: "hipsolver"
doc_link: "hipsolver:reference/precision"
data_types:
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "hipSPARSE"
tag: "hipsparse"
doc_link: "hipsparse:reference/precision"
data_types:
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "hipSPARSELt"
tag: "hipsparselt"
doc_link: "hipsparselt:reference/data-type-support"
data_types:
- type: "int8"
support: "✅"
- type: "float8 (E4M3)"
support: "✅"
- type: "float8 (E5M2)"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- name: "rocBLAS"
tag: "rocblas"
doc_link: "rocblas:reference/data-type-support"
data_types:
- type: "float16"
support: "⚠️"
- type: "bfloat16"
support: "⚠️"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "rocFFT"
tag: "rocfft"
doc_link: "rocfft:reference/api"
data_types:
- type: "float16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "rocRAND"
tag: "rocrand"
doc_link: "rocrand:api-reference/data-type-support"
data_types:
- type: "int8"
support: "Output only"
- type: "int16"
support: "Output only"
- type: "int32"
support: "Output only"
- type: "int64"
support: "Output only"
- type: "float16"
support: "Output only"
- type: "float32"
support: "Output only"
- type: "float64"
support: "Output only"
- name: "rocSOLVER"
tag: "rocsolver"
doc_link: "rocsolver:reference/precision"
data_types:
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "rocSPARSE"
tag: "rocsparse"
doc_link: "rocsparse:reference/precision"
data_types:
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "rocWMMA"
tag: "rocwmma"
doc_link: "rocwmma:api-reference/api-reference-guide"
data_types:
- type: "int8"
support: "✅"
- type: "int32"
support: "Output only"
- type: "float8 (E4M3)"
support: "Input only"
- type: "float8 (E5M2)"
support: "Input only"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "tensorfloat32"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "Tensile"
tag: "tensile"
doc_link: "tensile:reference/precision-support"
data_types:
- type: "int8"
support: "✅"
- type: "int32"
support: "✅"
- type: "float8 (E4M3)"
support: "✅"
- type: "float8 (E5M2)"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "tensorfloat32"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- group: "Primitives"
tag: "primitives"
libraries:
- name: "hipCUB"
tag: "hipcub"
doc_link: "hipcub:api-reference/data-type-support"
data_types:
- type: "int8"
support: "✅"
- type: "int16"
support: "✅"
- type: "int32"
support: "✅"
- type: "int64"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "hipTensor"
tag: "hiptensor"
doc_link: "hiptensor:api-reference/api-reference"
data_types:
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "rocPRIM"
tag: "rocprim"
doc_link: "rocprim:reference/data-type-support"
data_types:
- type: "int8"
support: "✅"
- type: "int16"
support: "✅"
- type: "int32"
support: "✅"
- type: "int64"
support: "✅"
- type: "float16"
support: "✅"
- type: "bfloat16"
support: "✅"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"
- name: "rocThrust"
tag: "rocthrust"
doc_link: "rocthrust:data-type-support"
data_types:
- type: "int8"
support: "✅"
- type: "int16"
support: "✅"
- type: "int32"
support: "✅"
- type: "int64"
support: "✅"
- type: "float16"
support: "⚠️"
- type: "bfloat16"
support: "⚠️"
- type: "float32"
support: "✅"
- type: "float64"
support: "✅"

2
docs/index.md
View File

@@ -65,7 +65,7 @@ ROCm documentation is organized into the following categories:
* [ROCm libraries](./reference/api-libraries.md)
* [ROCm tools, compilers, and runtimes](./reference/rocm-tools.md)
* [Accelerator and GPU hardware specifications](./reference/gpu-arch-specs.rst)
* [Precision support](./reference/precision-support.rst)
* [Data types and precision support](./reference/precision-support.rst)
* [Graph safe support](./reference/graph-safe-support.rst)
<!-- markdownlint-enable MD051 -->
:::

813
docs/reference/precision-support.rst
View File

File diff suppressed because it is too large Load Diff

2
docs/sphinx/_toc.yml.in
View File

@@ -180,7 +180,7 @@ subtrees:
- file: reference/gpu-arch-specs.rst
- file: reference/gpu-atomics-operation.rst
- file: reference/precision-support.rst
title: Precision support
title: Data types and precision support
- file: reference/graph-safe-support.rst
title: Graph safe support